Consumer LLM products expose the same underlying models through different surfaces. Each surface changes what enters the context window, what the model may call, and what the operator must verify.
What you will be able to do
2-hour session
- State, for any prompt, what is actually inside the context window: the question, the injected sources, the attached files, and the persistent memory.
- Predict which failure mode a given surface produces — stale recall, unfaithful summarization, or an unverified computation.
- Choose the cheapest surface that can answer a question correctly, instead of defaulting to the most capable one.
- Verify a retrieval-backed answer by opening its citations and labelling each claim supported, unsupported, or unverifiable.
- Read an emitted program as the reviewable artifact of a computation rather than trusting the printed number.
- Estimate the relative cost of a long-context turn and explain why pasting an entire document is rarely free.
- Audit persistent memory and custom instructions as a privacy surface, not merely a convenience feature.
Surfaces are context policies
A plain chat turn is a self-contained token stream with a knowledge cutoff. A search surface injects retrieved passages. A code surface injects repository files. A document surface injects an uploaded corpus. Choosing a surface is choosing what occupies the context window, not merely which button to press.
Retrieval shifts the failure mode
Without retrieval the risk is stale or fabricated recall. With retrieval the risk moves to source quality, passage selection, and unfaithful summarization. Retrieval reduces staleness; it does not remove the need to open the cited source and confirm the claim.
Delegated computation and code execution
When a surface can execute code, arithmetic and data work leave the token stream and run in an interpreter. This removes mental-arithmetic errors but introduces a new checkable artifact: the emitted program. Read the code, not only the number it printed. Extrapolations beyond the observed data range remain assumptions.
Operating cost and persistence
Longer context costs more and decodes more slowly, so bound what you paste. Memory, custom instructions, and saved configurations persist state across sessions: they are convenience features and a privacy surface, and they should be reviewed rather than accumulated silently.
Open the interactive — context budget Open the interactive — tokenizer
Framing informed by Andrej Karpathy, “How I use LLMs” (public lecture). Concepts here are restated and adapted for this course; the original is the authoritative source. https://www.youtube.com/watch?v=EWvNQjAaOHw
Quick check
Does enabling a search tool make an answer trustworthy?
Show answer
No. It changes the failure mode from stale recall to source and summarization quality. The cited source still has to be opened and checked.
Why prefer a code-executing surface for arithmetic?
Show answer
The computation runs in an interpreter instead of being predicted token by token, and it leaves an inspectable program as the artifact to review.
What is the cost of pasting a very long document into context?
Show answer
Higher price per turn and slower decoding, since attention and decoding scale with the tokens held in the window.
Practise this
Work the three labs, then check yourself against the 10-question quiz.